| # case-8b-rl-noopsd (step 1912) |
|
|
| **Clinical Agents for Seeking Evidence** - agentic harness learning for |
| longitudinal medical imaging. The model autonomously plans and calls tools inside a |
| multi-turn harness (model -> `qwen3_xml` tool call -> environment executes -> |
| observation -> ... -> final answer); it never answers single-shot. Three role |
| personas: **Anatomist** (organ segmentation / volumetry), **Radiologist** |
| (cross-visit image comparison), **Consultant** (report-grounded synthesis). |
| Paper: *From Given to Gathered Evidence: Agentic Learning for Longitudinal Medical Reasoning*. |
|
|
| Ablation: identical to the final model but **without the OPSD term** |
| (bucket-IV causal-KL) in the agentic RL objective. |
|
|
| - **Training**: agentic RL - GRPO + rubric-judge reward, causal-KL/OPSD disabled |
| - **Step**: 1912 |
| - **Base model**: Qwen3-VL-8B-Instruct |
| - **Format**: merged HF checkpoint (single `model.safetensors` + tokenizer/processor) |
| - **Serving tip**: run with a tool backend (the released mock-tool replay configs, |
| or the full LangGraph agents) and start vLLM with |
| `--enable-auto-tool-choice --tool-call-parser qwen3_xml` |
| (add `--enforce-eager` when the environment has no full CUDA toolkit) |
| - **Data**: trained on UK Biobank-derived, access-controlled data; no UKB-derived |
| data is redistributed here. Research use only. |
|
|
